Key Signals

• Cursor is in talks to raise at least $2 billion at a $50 billion valuation, nearly doubling its November 2025 figure. Thrive Capital and Andreessen Horowitz are expected to co-lead the oversubscribed round, with Battery Ventures and Nvidia also participating. Cursor forecasts ending 2026 at $6 billion+ ARR — tripling in 10 months — and has achieved slight gross margin profitability through its proprietary Composer model and lower-cost external models like Kimi. The company now faces its main competitive threat from Anthropic's Claude Code, making the race to build a durable enterprise platform the defining contest in AI-assisted development. [1]

• GitHub Copilot CLI v1.0.32 ships auto model selection, dynamically routing between GPT-5.4, GPT-5.3-Codex, Sonnet 4.6, and Haiku 4.5. The system chooses the most efficient model per session while offering all paid subscribers a 10% discount on the premium request multiplier, lowering effective costs for heavy users. This release also introduces document file attachments for agent reasoning, a --connect flag for remote sessions, and usage limit warnings at 75% and 90% thresholds — collectively making the CLI a more autonomous and cost-aware terminal agent. [2][3]

• Multiple analytics firms report that "tokenmaxxing" — maximizing AI token budgets — is producing volume without proportional value. Engineering managers see 80–90% code acceptance rates from tools like Claude Code, Cursor, and Codex, but post-merge churn drives real-world retention down to 10–30%. GitClear found AI users average 9.4x higher code churn; Faros AI measured an 861% increase under high AI adoption; Jellyfish data shows engineers achieve 2x throughput at 10x the token cost. The data suggests the industry urgently needs better metrics for AI-assisted development productivity beyond raw output. [4]

• OpenAI shed three senior executives in a single day as it folds its science and video teams into Codex. Former CPO Kevin Weil, Sora head Bill Peebles, and enterprise CTO Srinivas Narayanan all departed as OpenAI sunsets the Prism science workspace and reassigns its team under Codex lead Thibault Sottiaux. The restructuring signals that Codex is becoming OpenAI's "everything app," consolidating coding, science, and enterprise workflows into a single desktop application — a sharp pivot toward monetizable developer tools ahead of a planned IPO. [5]

• Anthropic CEO Dario Amodei met White House Chief of Staff Susie Wiles in what both sides called "productive and constructive" talks over access to Mythos. The frontier model discovered thousands of zero-day vulnerabilities across every major OS and browser with 83% first-attempt exploit success during testing, prompting Anthropic to restrict it to ~40 vetted organizations through Project Glasswing rather than releasing it publicly. Treasury, CISA, and parts of the intelligence community are already testing Mythos Preview, creating a paradox where the government that blacklisted Anthropic over safety restrictions now needs its most powerful model for cyber defense. [6][7]

AI Coding News

• Meta reports a 4x improvement in bug detection using Just-in-Time testing that generates targeted tests during code review instead of relying on static test suites. The system uses LLMs, mutation testing, and an intent-aware "Dodgy Diff" architecture that reframes each code change as a semantic signal, inferring developer intent and injecting synthetic defects to validate test effectiveness. Evaluated on over 22,000 generated tests, the approach detected up to 20x more meaningful failures than coincidental matches, with 8 confirmed real defects out of 41 issues surfaced in one evaluation subset. As AI-generated code accelerates pull request volume, JiT testing represents a critical shift from static correctness validation toward change-specific fault detection that scales with agentic workflows. [8]

• AWS released Agent Registry in preview as part of Amazon Bedrock AgentCore, providing a centralized catalog for discovering, governing, and reusing AI agents across organizations. The registry indexes agents regardless of where they run and supports both MCP and A2A protocols natively, addressing the growing challenge of "agent sprawl" as enterprises deploy hundreds of autonomous AI agents across teams. Microsoft, Google Cloud, and the ACP Registry offer competing solutions, signaling that agent governance infrastructure is rapidly becoming a critical enterprise requirement. [9]

Feature Update

• Claude Code v2.1.113 makes a major architectural shift to native binary execution and hardens sandbox security. The CLI now spawns a platform-specific native binary instead of bundled JavaScript, a change that should improve startup performance and reduce dependency on Node.js. Three new security measures close loopholes in Bash deny rules: exec wrappers like env/sudo/watch/setsid are now matched, find -exec/-delete is no longer auto-approved under Bash allow rules, and macOS /private/{etc,var,tmp,home} paths are treated as dangerous removal targets. The release also improves /ultrareview with parallelized checks and fixes MCP concurrent-call timeout handling. [10]

• Anthropic introduced a multi-agent Code Review system for Claude Code that dispatches parallel AI reviewers to analyze pull requests. In internal use across most of Anthropic's own PRs over several months, substantive review comments increased from 16% to 54% of pull requests. On PRs exceeding 1,000 lines, 84% generated findings averaging 7.5 issues, with fewer than 1% of findings marked incorrect by engineers. The feature costs approximately $15–25 per pull request at current Opus pricing and is available in research preview for Team and Enterprise users, competing directly with GitHub Copilot code review and CodeRabbit. [11]

• Anthropic launched Claude Design in research preview, a visual design tool powered by Claude Opus 4.7 for generating prototypes, slides, one-pagers, and design systems. Users can feed their codebase and design files to create a custom design system, then generate artifacts with real-time tweaking via sliders and direct element editing. Exports go to PDF, PPTX, HTML, or Canva for collaborative refinement — notably not Figma, whose stock dropped 5% at the announcement following CPO Mike Krieger's departure from Figma's board. The tool is token-intensive; building one design system and prototype consumed over 50% of a weekly Pro allotment during early testing. [12][13]

• OpenAI Codex published four alpha builds (0.122.0-alpha.6 through alpha.9) on April 17, continuing rapid iteration on the Rust-based CLI. These incremental releases follow the stable 0.122.0 release from April 16, which added computer use, in-app browsing, image generation, memory, and plugins to the Codex desktop app for macOS and Windows. The pace of daily alpha drops reflects the aggressive development cadence as OpenAI positions Codex as its unified developer platform. [14]

• OpenCode shipped three releases in a single day: v1.4.8, v1.4.9, and v1.4.10, adding LLM Gateway as a provider and Claude Opus 4.7 compatibility. v1.4.9 added LLM Gateway support with config and model usage reporting while restricting GitHub Copilot Opus 4.7 models to medium reasoning effort to avoid unsupported variants. v1.4.8 enabled Azure prompt caching with a default per-session cache key and fixed a crash in experimental mode. v1.4.10 restored workspace history on connect and normalized provider metadata defaults for incomplete catalog data. [15]

• Gemini CLI v0.38.2 shipped as a patch release cherry-picking a fix from v0.38.1. This is a minor maintenance release for the Google-maintained terminal AI agent, addressing a specific bug without broader feature changes. [16]

• Kiro added support for Claude Opus 4.7 with a 1M token context window and a 2.2x credit multiplier. The model offers stronger agentic coding performance, more precise instruction following, and 3x higher resolution vision compared to prior Opus versions. The rollout is initially limited to Pro, Pro+, and Power tier subscribers using AWS IAM Identity Center sign-in, with inference in us-east-1 and eu-central-1 regions. [17]